Physics

From PRIMO.ai
Jump to navigation Jump to search

Youtube search... ...Google search

Physics and artificial intelligence have become partners in both directions. Physics gave AI some of its oldest and deepest ideas: the energy landscapes of Hopfield networks, the statistical mechanics behind Boltzmann machines, and the thermodynamics that inspired diffusion models. In return, AI now gives physics new instruments: neural networks that emulate expensive simulations thousands of times faster, controllers that steer fusion plasmas, models that forecast the weather, and algorithms that sift detector data for signals humans would miss. This page covers both directions, from the physics inside machine learning to the machine learning inside modern physics, and it pays special attention to the hardest question in the field: when a model predicts nature well, has it learned physics or only a convincing imitation of it?

The difference from conventional approaches is easiest to see in how each one handles knowledge. Classical physics starts from principles, such as conservation of energy or Newton's laws, and derives equations that are then solved by hand or by numerical simulation. Pure machine learning starts from data and learns a function with almost no built-in assumptions, which makes it flexible but data-hungry and prone to physically impossible answers. Physics-informed machine learning sits between the two, building known laws, symmetries, and conservation principles into the data, the architecture, or the loss function, so that models need less data and extrapolate more safely. A fourth approach, scientific discovery with AI, runs the process in reverse: it trains a model on data and then tries to extract a compact, human-readable law from what the model learned, which was the goal of Tailin Wu and Max Tegmark's "AI Physicist" and remains the goal of symbolic regression.

The relationship has a long history and, recently, official recognition. John Hopfield's 1982 network borrowed the mathematics of magnetic spin systems, and the Boltzmann machine that Geoffrey Hinton developed with Terrence Sejnowski in 1985 borrowed Ludwig Boltzmann's statistical mechanics; in 2024 Hopfield and Hinton shared the Nobel Prize in Physics, while the creators of AlphaFold shared the chemistry prize with David Baker. The award was not free of dispute, since Jürgen Schmidhuber argued that earlier work by researchers such as Shun-ichi Amari and Alexey Ivakhnenko deserved more credit, and some physicists questioned whether neural networks belonged in a physics prize at all. The 2025 physics prize went to John Clarke, Michel Devoret, and John Martinis for macroscopic quantum effects in electrical circuits, the physics underlying today's superconducting quantum processors, and the 2026 prize went to Francis Halzen for the IceCube Neutrino Observatory, whose 2023 discovery of neutrinos from the Milky Way's Galactic plane depended on deep learning to pick signals out of the background. Between those milestones, AI weather models began beating physics-based forecasts, reinforcement learning learned to shape fusion plasmas, and in 2025 and 2026 AI systems reached gold-medal level on the theory exam of the International Physics Olympiad and claimed results on famous problems in fluid dynamics.

Physics also sets hard limits on AI itself. Every operation a model performs costs energy, moving data costs far more than computing on it, and today's chips use roughly a hundred million times more energy per operation than the thermodynamic floor set by Landauer's principle; those numbers are worked out in the Physics section of the Mathematics page. For the growing role of AI in proving theorems about the equations of physics, including the 2026 claims about the Navier-Stokes equations, see What AI Has Given Mathematics.


What I cannot create, I do not understand.
Written on Richard Feynman's Caltech blackboard at the time of his death in 1988.
For AI in physics the lesson cuts both ways: a model that can generate nature has not yet explained it, and a law that no one can compute with is of limited use.


Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube

The Mechanism

Steve Brunton of the University of Washington describes machine learning as five stages, and physics can enter at every one of them. The list below follows that framework and adds a sixth stage, extraction, in which a trained model is turned back into physics a person can read. Most successful systems combine several stages rather than relying on just one.

  1. Choose what to model. Decide whether to learn a whole system, a correction to a known model (a "closure" term for turbulence, for example), an operator that maps one function to another, or a controller. Learning only the part that physics cannot already describe is usually far more data-efficient than learning everything.
  2. Curate the data. Training data come from experiments, observatories, or simulations. Known symmetries can be added by data augmentation, for example by rotating examples of a rotation-invariant system, and datasets such as Polymathic AI's "The Well" gather many kinds of simulation in one place.
  3. Build physics into the architecture. Graph neural networks treat particles or mesh points as nodes and their interactions as edges; equivariant networks give consistent answers when the input is rotated or shifted; Hamiltonian and Lagrangian neural networks conserve energy by construction; and a 2026 EPFL model, Dynami-CAL GraphNet, builds Newton's third law into its structure so that every predicted force has an equal and opposite partner.
  4. Put physics into the loss. Physics-informed neural networks (PINNs) add a penalty whenever the network's output violates a governing equation: L = Ldata + λ · mean(|𝒩[uθ](x, t)|2), where 𝒩 is the differential operator, uθ is the network's candidate solution, and λ balances fitting the data against obeying the physics.
  5. Optimize with differentiable physics. When a simulator is written in an automatic-differentiation framework, gradients flow through the physics itself, so a model can be trained, or a design optimized, end to end. Google DeepMind's TORAX plasma simulator, written in JAX, is an example of this differentiable programming approach.
  6. Extract the law. Interpretation methods turn a trained network into equations: sparse regression (SINDy), symbolic regression tools such as AI Feynman and PySR, or Miles Cranmer's strategy of training a network with strong built-in assumptions and then distilling its learned pieces into formulas.

Dependencies and conditions. These methods work best when the governing equations are at least partly known, when the data cover the conditions the model will face, and when an independent check, such as a high-fidelity simulation or an experiment, exists. Without such checks, a model can fit its training data beautifully while violating conservation laws or failing badly outside the training range, a form of out-of-distribution failure that is especially dangerous in safety-critical engineering. Physics constraints reduce this risk but do not remove it, because a loss penalty can be traded off against data fit and an architecture can encode the wrong symmetry.

The Physics

Many core ideas in AI are physics under a different name. In a Hopfield network, each neuron behaves like a magnetic spin pointing up or down, and the network has an energy E = -½ Σij wij si sj, the same form as the Ising model of magnetism; stored memories are valleys in this energy landscape, and recalling a memory means letting the system roll downhill. Hinton and Sejnowski's Boltzmann machine made the network random, so that each configuration appears with probability proportional to e-E/T, the Boltzmann distribution of statistical mechanics, and learning means reshaping the landscape so that realistic data sit at low energy. This family of methods is now called energy-based models. The same distribution survives inside every modern language model: the softmax function turns scores into probabilities exactly as the Boltzmann distribution turns energies into probabilities, and the sampling "temperature" is borrowed directly from thermodynamics.

Generative AI drew on physics again in 2015, when Jascha Sohl-Dickstein and colleagues used ideas from non-equilibrium thermodynamics to build what became diffusion models: data are gradually dissolved into noise, the way ink spreads through water, and a network learns to run the process backward. Sampling from these models resembles Langevin dynamics, the jittery motion of a particle in a heat bath, and the connection is more than a metaphor because the same stochastic differential equations describe both. Physicists have also argued that deep learning works because the world it models is physically simple. Henry Lin, Max Tegmark, and David Rolnick noted that physical laws tend to be low-order, local, symmetric, and hierarchical, which lets relatively small networks represent them, and Pankaj Mehta and David Schwab showed that a form of the renormalization group, physics' method for zooming out from microscopic to large-scale behavior, maps exactly onto deep networks built from restricted Boltzmann machines.

Physics may also supply the missing theory of deep learning. Neural scaling laws, in which loss falls as a power of model size, data, and compute, resemble the power laws of critical phenomena, and researchers use statistical mechanics, random matrix theory, and dynamical systems to explain why heavily overparameterized networks generalize instead of memorizing, and why some abilities appear to arrive suddenly in a way that looks like emergence. These explanations are still partial and contested. The scaling laws themselves remain empirical fits, and the analogy to phase transitions is best treated as a guide for research rather than an established result.

Finally, physics constrains how AI can be built. Training and serving frontier models consumes gigawatts of energy, which is why researchers pursue neuromorphic, analog, optical, and thermodynamic hardware that computes with physical processes directly, and why the cost of moving bits dominates chip design. Quantum computing is the most radical version of this idea, using superposition and interference as computational resources; its progress rests on the superconducting circuits recognized by the 2025 Nobel Prize and, as described below, increasingly on AI decoders that correct its errors. See processing units and Quantum for the hardware landscape.

Why This Matters

  • Scientific importance. Physics produces some of the largest datasets in science, from billions of collisions at the Large Hadron Collider to sky surveys of hundreds of millions of objects, and much of that data can no longer be analyzed without machine learning. AI also offers a new route to discovery, finding structure in complex many-body systems, such as dusty plasmas, where analytic theory struggles.
  • Engineering implications. Fast surrogate models can replace hours of simulation with seconds of inference, which changes how aircraft, reactors, batteries, and chips are designed. The catch is that a surrogate is only trustworthy inside the conditions it was trained on, so engineers increasingly pair AI with physics-based digital twins and conventional solvers for verification.
  • Practical consequences. Weather forecasts on phones, fusion reactor control, materials discovery, and the hardware triggers that decide which collider events to keep already depend on AI. Practitioners need to test these models against strong baselines and physical constraints, not only on held-out data, using the evaluation practices described in model evaluation.
  • Conceptual implications. A model that predicts well has not necessarily explained anything. Physics has always prized compact laws that generalize, and the push for interpretable and symbolic models reflects that value; a 2022 Columbia University study, in which an algorithm watching videos of physical systems settled on descriptions that did not obviously match the familiar variables, showed that machine-discovered physics may look different from human physics even when both work.
  • Limitations and uncertainty. A 2024 analysis in Nature Machine Intelligence by Nick McGreivy and Ammar Hakim found that much published work on machine learning for fluid equations compared against weak baselines, inflating claimed speedups. PINNs can fail on stiff, chaotic, or high-frequency problems, pretrained models can miss genuinely new physics by mapping it onto familiar patterns, and benchmark scores for AI physics reasoning have proven fragile in both directions.
  • What would change if current trends hold. Foundation models pretrained across many physical systems could become standard starting points, as language models already are for text, and AI agents could run simulation and analysis campaigns largely on their own. The bottleneck would then move, as it already has in mathematics, from producing results to verifying and understanding them.

Recent Developments

AI-for-physics research moved from demonstrations to deployment between 2022 and 2026, and from narrow models trained for one task toward general models trained across many. The table summarizes the milestones and how solid each one is, ranging from peer-reviewed and operational systems to self-reported results still under review. Domain-by-domain details appear in AI Across the Physical Sciences.

Date Development What was shown Evidence status
Feb 2022 Deep reinforcement learning controls a tokamak (DeepMind and EPFL) One neural network drove all 19 magnetic coils of the TCV tokamak and sculpted several plasma shapes Peer reviewed (Nature)
Jul 2022 "Alternative physics" (Columbia University) An algorithm estimated how many variables describe physical systems from video, without prior knowledge Peer reviewed; interpretation debated
Jun 2023 Neutrinos from the Milky Way (IceCube) Deep-learning event selection kept more than 20 times as many events and revealed Galactic neutrinos at 4.5σ Peer reviewed (Science)
Nov 2023 GraphCast (Google DeepMind) Ten-day global forecasts more accurate than the leading physics-based system on about 90 percent of 1,380 targets Peer reviewed (Science)
Oct 2024 Nobel Prize to Hopfield and Hinton Physics prize for foundational neural network research Awarded; priority disputed by some
Nov 2024 AlphaQubit (Google DeepMind) A neural network decoder outperformed leading decoders on quantum error-correction data from Google hardware Peer reviewed (Nature)
Feb 2025 ECMWF AIFS becomes operational A machine-learning forecast model enters routine use at Europe's leading weather center Operational
2025 Dusty plasma forces (Emory University) A physics-tailored network learned non-reciprocal forces between particles and corrected theoretical assumptions Peer reviewed
Sep 2025 Physics Supernova An agent matched gold medalists on IPhO 2025 theory problems, ranking 14th among 406 contestants Preprint; graded by the authors
Sep 2025 CritPt benchmark 71 unpublished, research-level physics challenges; the best models initially solved under 10 percent Preprint and public leaderboard
Oct 2025 Google DeepMind and Commonwealth Fusion Systems TORAX simulator and AI control strategies for the SPARC tokamak Partnership; results pending
Nov 2025 Walrus and AION-1 (Polymathic AI) Cross-domain foundation models for continuum dynamics and for astronomy Preprint and NeurIPS; open weights
Feb 2026 Dynami-CAL GraphNet (EPFL) Newton's third law built into a graph network for stable, physically consistent simulation Peer reviewed (Nature Communications)
Mar 2026 AI rediscovers hadron patterns (NYU Abu Dhabi) Organizing principles of the quark model and Regge trajectories recovered from particle data Peer reviewed (JHEP)
Jul 2026 IPhO 2026 AI systems scored 28.6 of 30 on the theory exam under official grading Officially graded; theory exam only
Sep 2026 Navier-Stokes blow-up claim (OpenAI) Finite-time singularity for the forced 3D Navier-Stokes equations, with a Lean proof Under expert review; unforced case open
Sep 2026 WeatherNext 3 (Google) Hourly global forecasts at up to 5 km resolution, trained to use raw satellite observations Deployed; described by Google as experimental
Sep 2026 Physics benchmark re-grading An expert audit argued that low scores on leading physics benchmarks largely reflected flawed grading Preprint; contested
Oct 2026 Nobel Prize to Francis Halzen IceCube and the discovery of astrophysical neutrinos Awarded

AI That Finds Physical Laws

The strongest evidence that AI can do more than fit data comes from a handful of studies in which a model found something physicists did not already know. In a 2025 study, Emory University physicists Justin Burton and Ilya Nemenman built a neural network that respected the symmetries of a dusty plasma, a cloud of charged dust grains suspended in ionized gas, and trained it on three-dimensional particle trajectories from experiments. The model learned the effective forces between particles, including non-reciprocal forces in which one particle pushes on another harder than it is pushed back, and it showed that these forces depend on particle size in a way existing theory had assumed away. Nemenman has noted how rare such cases are, observing that very few discoveries in science have been made directly by an AI system rather than with its help.

Other recent work tests whether AI could have made famous discoveries on its own. In a March 2026 study in the Journal of High Energy Physics, NYU Abu Dhabi researchers showed that relatively simple machine learning methods, given only particle data, recovered the organizing patterns behind dozens of hadrons that physicists took decades to explain with quarks, along with Regge trajectories linking particle mass to spin. Multi-agent systems built on large language models now combine literature search, hypothesis generation, and symbolic regression to propose physical laws, and benchmarks such as NewtonBench test them on hundreds of law-discovery tasks. A June 2026 cosmology study offers a caution: transfer learning from standard-model simulations made inference faster, but it could also cause a model to mistake signs of new physics, such as evolving dark energy, for familiar patterns it had already learned.

Foundation Models for Physics

Language models showed that one large model pretrained on broad data can be adapted to many tasks, and physicists are now testing whether the same is true for physical systems. The Polymathic AI collaboration, which includes the Flatiron Institute, the University of Cambridge, and New York University, released Walrus in November 2025: a 1.3-billion-parameter Transformer pretrained on 19 scenarios spanning astrophysics, geoscience, rheology, plasma physics, acoustics, and classical fluids, drawn from the team's dataset "The Well." The authors report state-of-the-art results on 56 of 65 metrics across 26 continuum simulation tasks, and the open weights on Hugging Face let researchers fine-tune the model on new problems from a short window of simulation data. A sister model, AION-1, does the same for astronomy, pretraining on more than 200 million objects across 39 kinds of measurements from five major sky surveys.

These foundation models are promising but young. They are trained mostly on simulations, so they inherit the assumptions and numerical errors of those simulations, and long predictions can drift as small errors compound, particularly in chaotic systems. They also face the cosmology caution above: a model that has seen a great deal of standard physics may be slow to recognize something genuinely new. Miles Cranmer, one of the project's leaders, argues that the larger goal is not only better emulators but models whose internal representations can be interpreted to reveal physics, a theme of his 2024 lecture listed under Additional Viewing.

Reasoning Models Take On Physics

Reasoning models have followed the same path in physics that they followed in mathematics. In 2025, the Physics Supernova agent, built around a general LLM with tools for reading figures and checking answers, scored at gold-medal level on the theory problems of the International Physics Olympiad, ranking 14th among 406 human contestants, although the grading was done by the authors. At IPhO 2026 in Bucaramanga, Colombia, AI entries sat the theory exam under official conditions: Huawei's assistant Celia and a model from the startup Asteromorph each reported a score of 28.6 out of 30, above the gold threshold, and the IPhO secretariat called Celia's performance "a historic milestone." Meta reported a perfect theory score at the 2026 Asian Physics Olympiad without disclosing which model it used. None of these systems attempted the experimental exam, which tests the hands-on side of physics that text models cannot reach.

Research-level physics is a harder test, and the evidence is genuinely mixed. CritPt, released in September 2025, contains 71 unpublished challenges across eleven fields, written by more than 50 working physicists to resemble the first project of a capable graduate student. The best models initially solved under 10 percent, and by spring 2026 the leading scores had climbed to about 30 percent on the independent leaderboard. Then a September 2026 audit by expert re-graders argued that many failures came from broken answer keys and grading rules rather than from the models; after correction, it reported that one frontier model's pass rate on 54 retained CritPt challenges reached about 94 percent, and it concluded that closed-ended, text-only physics problems are close to saturation. That conclusion is contested, and it applies only to problems with checkable final answers. The episode repeats a lesson from the statistics of AI evaluation: a benchmark is an experiment, and both its questions and its graders can be wrong.

Open Questions

  • Can machine learning discover genuinely new physical laws, rather than rediscovering known ones or fitting effective forces in a single system?
  • Will physics foundation models generalize to regimes their training simulations never covered, and how will anyone know when they have not?
  • Can AI-designed controllers be certified safe enough for power plants, aircraft, and fusion reactors?
  • Will learned weather models stay reliable as the climate moves outside the historical record they were trained on?
  • How should physics benchmarks be built and graded so that scores mean what they claim?
  • Will the unforced Navier-Stokes problem fall, and would a blow-up proof change how turbulence is understood or simulated?

Additional Viewing

Load video
YouTube

2024 Nobel Prize lectures in physics | John Hopfield and Geoffrey Hinton
The official Nobel lectures from Stockholm University's Aula Magna in December 2024, in which John Hopfield presents "Physics is a point of view" and Geoffrey Hinton explains Boltzmann machines to a general audience without equations. Together they are a first-hand account of the spin-glass and statistical-mechanics roots of neural networks described in The Physics, from energy landscapes to the restricted Boltzmann machines that helped launch deep learning.
Watch on YouTube

Load video
YouTube

Physics Informed Machine Learning: High Level Overview of AI and ML in Science and Engineering | Steve Brunton
University of Washington professor Steve Brunton's roughly 47-minute introduction breaks machine learning into five stages (problem, data, architecture, loss, and optimization) and shows how physics can be embedded at each one, using a pendulum as a running case study. It is the lecture behind the framework in The Mechanism and the best entry point to the methods in Physics-Informed AI, and it opens a longer free course on the same topic.
Watch on YouTube

Load video
YouTube

Miles Cranmer: The Next Great Scientific Theory is Hiding Inside a Neural Network | Simons Foundation
In this April 2024 Simons Foundation Presidential Lecture, Cambridge astrophysicist and machine learning researcher Miles Cranmer argues that scientists should train neural networks on data and then interpret what the networks learned, distilling them into equations with symbolic regression. He also introduces the Polymathic AI initiative behind Walrus and AION-1, making the talk a direct companion to Equation Discovery and Symbolic Regression and Foundation Models for Physics.
Watch on YouTube

Load video
YouTube

AI Crisis in Math: Navier-Stokes, Riemann Hypothesis & Complexity Theory | Lex Fridman Podcast #503
Released October 26, 2026, this conversation includes long chapters with fluid dynamicist Steven Brunton on why the Navier-Stokes equations are so hard, what OpenAI's claimed blow-up result does and does not mean for turbulence and engineering design, and how nature computes. It is the most current discussion of the material in Fluids, Turbulence, and Mathematical Physics, alongside mathematicians Alex Kontorovich and Ankur Moitra on the broader AI-and-mathematics debate.
Watch on YouTube

Physics-Informed AI

Physics-informed AI is the practice of building known physics into machine learning models so that they need less data, respect conservation laws, and extrapolate more reliably. The videos below include a short explainer of how physics-informed neural networks work and Max Tegmark's MIT talk on the deeper connections between physics and deep learning. The subsections that follow explain the main families of methods and where each one tends to succeed or fail.

Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube

Physics-Informed Neural Networks (PINNs)

A physics-informed neural network represents the solution of a physical problem, such as the temperature in a metal plate or the velocity of a fluid, as a neural network that takes position and time as inputs. Training pushes the network to match whatever measurements exist while also satisfying the governing differential equation at many sample points, using automatic differentiation to compute the derivatives the equation requires. Because the physics acts as a strong regularizer, PINNs can work with sparse, noisy data, and they are especially attractive for inverse problems, where the goal is to infer hidden quantities such as material properties or a source term from limited observations. They also need no simulation mesh, which helps in awkward geometries.

PINNs have well-documented weaknesses. Training can stall because the data-fit and physics terms pull in different directions, networks learn smooth, low-frequency features first and struggle with sharp fronts or turbulence, and stiff or chaotic equations often defeat them. For ordinary forward problems where the equations and conditions are fully known, a mature numerical solver is usually faster and more accurate, which is part of why McGreivy and Hakim's review of weak baselines matters. Current research addresses these issues with adaptive weighting of the loss terms, domain decomposition, curriculum training, and hybrid schemes that let a conventional solver handle what it does well while a network learns only the missing physics.

Neural Operators and Surrogate Models

Where a PINN learns one solution, a neural operator learns the whole mapping from inputs to solutions, for example from an initial flow field to the flow a second later, so that a single trained model can answer many related questions in milliseconds. The Fourier Neural Operator does this by acting on the low-frequency Fourier modes of a field, an idea explained in the Fourier section of the Mathematics page, while DeepONet and graph-based operators handle irregular geometries. The same idea, scaled up with graph networks and Transformers, underlies AI weather models and the Walrus foundation model. These surrogates are now used to explore design spaces, run large ensembles for uncertainty estimates, and accelerate inverse problems, with the conventional solver held in reserve to check the most important answers.

Symmetry, Conservation, and Geometry

Emmy Noether proved in 1918 that every continuous symmetry of a physical system corresponds to a conserved quantity: symmetry in time gives conservation of energy, symmetry in space gives conservation of momentum, and rotational symmetry gives conservation of angular momentum. Machine learning has taken that insight to heart. Equivariant networks guarantee that rotating or shifting the input rotates or shifts the output in the same way, Hamiltonian and Lagrangian neural networks learn an energy-like function and derive the motion from it so that energy is conserved by construction, and geometric deep learning treats molecules, particles, and meshes as graphs whose structure mirrors the physics. EPFL's 2026 Dynami-CAL GraphNet shows the payoff: by building Newton's third law into the network, it produced stable simulations of systems it had never seen, an extrapolation that unconstrained models rarely manage. Building in symmetry is one of the most reliable ways to make physics models data-efficient, but choosing the wrong symmetry is just as reliable a way to make them wrong.

Equation Discovery and Symbolic Regression

The oldest dream of AI in physics is a machine that looks at data and writes down the law. Symbolic regression searches over mathematical expressions, typically with evolutionary algorithms, for the simplest formula that fits the data. Brunton, Proctor, and Kutz's SINDy method (2016) assumes that most dynamics involve only a few terms and uses sparse regression to pick them from a large library of candidates. AI Feynman (2020) combined neural networks with physics-inspired tricks such as dimensional analysis and symmetry testing and recovered all 100 equations in a test set drawn from The Feynman Lectures on Physics, where earlier software had found 71. Cranmer's PySR has since become a standard open-source tool, and in 2020 he and colleagues used a graph network plus symbolic regression to find a new formula predicting the concentration of dark matter from the distribution of surrounding cosmic structures.

These methods connect machine learning to neuro-symbolic and symbolic AI, and to representation learning, since the hard part is often choosing the right variables before any equation can be found. The 2022 Columbia study, which estimated the number of hidden variables in systems such as a swinging double pendulum directly from video, showed both the promise and the puzzle: the algorithm's variables did not map neatly onto the angles and velocities a physicist would choose. The newest systems put AI agents in the loop, letting language models propose candidate variables and hypotheses from the literature while symbolic regression and simulation test them. Their record so far is mostly rediscovery, which is a necessary first step but not yet a new physics.

AI Across the Physical Sciences

Almost every branch of physics now uses machine learning, but the role it plays varies widely. In some fields AI is a forecasting engine, in others a controller, a filter for overwhelming data, or a search tool for new materials and mathematical structures. The sections below summarize the most important applications as of late 2026, with attention to what has been demonstrated, what is operational, and what remains a claim.

Weather and Climate

Weather forecasting is the clearest success story for AI in the physical sciences. Google DeepMind's GraphCast, published in Science in 2023, used a graph network trained on decades of reanalysis data to produce ten-day global forecasts in under a minute on a single chip, beating the European Centre for Medium-Range Weather Forecasts' flagship physics model on about 90 percent of 1,380 test targets. GenCast (2024) added probabilistic ensembles, ECMWF made its own machine-learning model, AIFS, operational in early 2025, and Google's WeatherNext 2 brought these forecasts into consumer products in November 2025. On September 3, 2026, Google released WeatherNext 3, which refreshes global forecasts hourly at up to 5 km resolution and learns from live geostationary satellite observations rather than only from the output of physics-based models, and which led the independent Operational WeatherBench comparison at launch.

Important caveats remain for meteorology and climate science. Most AI weather models still learn from reanalysis data produced by physics models, extreme events are rare in the training record, and Google itself labels WeatherNext as an experimental system that should not replace official warnings. Climate projection is a different problem from weather forecasting, because the future climate lies partly outside the historical data, which is why hybrid approaches such as Google's NeuralGCM, which pairs a differentiable physics core with learned components, are receiving attention for long-range simulation.

Fusion and Plasma Physics

A tokamak confines plasma hotter than the Sun's core with magnetic fields that must be adjusted thousands of times per second. In 2022, DeepMind and the Swiss Plasma Center at EPFL trained a single network with deep reinforcement learning in simulation and then used it to control all 19 magnetic coils of the real TCV tokamak, sculpting shapes from conventional ones to two separate plasma "droplets" at once. In 2024, a Princeton-led team showed on the DIII-D tokamak that an AI controller could forecast dangerous tearing instabilities up to about 300 milliseconds ahead and adjust the plasma to avoid them. In October 2025, Google DeepMind announced a partnership with Commonwealth Fusion Systems to use its differentiable, open-source TORAX simulator, combined with reinforcement learning and evolutionary search, to plan operations and develop an AI pilot for the SPARC tokamak, which the company expects to reach first plasma in late 2026 or early 2027.

AI is reshaping fusion research beyond control, from fast surrogate models of plasma turbulence to defect detection in magnet manufacturing. The open question is certification: a controller trained in simulation must be trusted on a machine where a disruption can damage hardware, so these systems are being introduced gradually, with conventional controllers and safety interlocks behind them. The Emory dusty plasma result described above shows the scientific side of the same story, with AI helping explain plasma behavior rather than only steering it.

Particle, Nuclear, and Astroparticle Physics

Particle physics adopted machine learning early because its data volumes demand it. At the Large Hadron Collider, neural networks run on FPGAs in the first-level triggers that decide within microseconds which collisions to keep, graph networks reconstruct particle tracks, and Transformer-based taggers identify the sprays of particles left by quarks and gluons far better than hand-built variables could. Model-independent searches use anomaly detection to look for any departure from the Standard Model rather than one specific hypothesis, and the 2026 NYU Abu Dhabi study showed that simple methods can recover the quark model's organizing patterns from data alone.

The 2026 Nobel Prize in Physics highlights the same trend in astroparticle physics. IceCube, a cubic kilometer of sensors buried in the Antarctic ice, sees mostly background from atmospheric muons, which outnumber astrophysical neutrinos by roughly a hundred million to one. For its 2023 observation of neutrinos from the Galactic plane, the collaboration replaced earlier event selection with convolutional neural networks and a hybrid of maximum likelihood and deep learning that builds the detector's symmetries into the network; the new selection kept more than 20 times as many events and revealed the Milky Way's neutrino glow at 4.5σ. Graph networks and Transformer models now push IceCube's reconstruction further, in some cases running fast enough to support real-time alerts.

Astronomy, Cosmology, and Gravitational Waves

Astronomy faces a flood of data from surveys that catalog hundreds of millions of objects, and machine learning now classifies galaxies, finds strong gravitational lenses, and flags transients for follow-up. Cosmology increasingly relies on simulation-based inference, in which neural networks learn the relationship between cosmological parameters and simulated observations and then perform Bayesian inference without an explicit likelihood formula. In gravitational-wave astronomy, neural posterior estimation (the DINGO family of methods) has analyzed binary neutron star signals in about a second, fast enough to point telescopes at a merger while its light is still fresh. Multimodal foundation models such as AION-1 aim to unify images, spectra, and time series in a single representation. The June 2026 transfer-learning study is a reminder that speed brings a risk: a model pretrained on the standard cosmological model may interpret evidence for massive neutrinos, modified gravity, or evolving dark energy as noise around what it already knows.

Quantum Physics

AI and quantum physics now support each other. In 2017, Giuseppe Carleo and Matthias Troyer showed that neural networks can represent the quantum states of many interacting particles, opening a new way to simulate materials and molecules that is now a standard tool. Algorithms have designed quantum optics experiments that human physicists had not considered, starting with Mario Krenn's MELVIN in 2016, and reinforcement learning tunes quantum devices and optimizes control pulses. On the hardware side, Google DeepMind's AlphaQubit (2024) used a Transformer-based neural decoder to correct errors in Google's superconducting quantum processors more accurately than leading conventional decoders, a key step because useful quantum computing depends on fast, accurate error correction. The superconducting circuits these decoders protect descend directly from the experiments recognized by the 2025 Nobel Prize in Physics.

Materials, Chemistry, and Condensed Matter

Machine-learned interatomic potentials now let researchers simulate millions of atoms with close to quantum-mechanical accuracy at a tiny fraction of the cost, transforming computational materials science and chemistry. In 2023, Google DeepMind's GNoME predicted about 2.2 million new crystal structures, roughly 380,000 of them stable, and in 2025 Microsoft's MatterGen showed that a generative model can propose materials with requested properties. Critics, including materials chemists Anthony Cheetham and Ram Seshadri, argued that many GNoME predictions were neither novel nor useful, a reminder that predicted stability is not the same as a synthesizable, valuable material. Related advances in molecular AI and protein structure prediction, recognized by the 2024 Nobel Prize in Chemistry, show how far physics-aware learning has already reached into neighboring sciences.

Fluids, Turbulence, and Mathematical Physics

The Navier-Stokes equations express Newton's second law for fluids, and they govern everything from airflow over aircraft wings to blood in artificial hearts. Turbulence remains one of the great unsolved problems of classical physics, and machine learning is used to build closure models for unresolved small eddies, to accelerate simulations, to reduce drag through learned flow control, and to discover reduced-order models of complex flows, an area where Steven Brunton's group has been influential. Brunton often cites the view, associated with Richard Feynman, that knowing a differential equation does not tell us what phenomena it contains, which is exactly the gap that simulation and learning try to fill.

AI has also entered the mathematics of fluids. In 2025, Google DeepMind and academic collaborators used physics-informed neural networks with computer-assisted proof techniques to discover new families of unstable self-similar singularities in equations related to the Euler equations. In September 2026, OpenAI claimed a Lean-verified proof that the three-dimensional Navier-Stokes equations, driven by a smooth external force, can develop a singularity in finite time, after Tristan Buckmaster and Levent Alpöge had announced related results for the Euler equations; the claim covers only the forced versions of the Clay Millennium Prize problem, the unforced versions remain open, and the Clay Mathematics Institute has not accepted it. Brunton has stressed that even a valid blow-up proof does not directly make turbulence easier to predict or aircraft easier to design, though it sharpens what the equations can and cannot do. The full story, including the credit dispute and the verification status, is told in What AI Has Given Mathematics.

Retrieved from "https://primo.ai/w/index.php?title=Physics&oldid=50047"